Novel View Synthesis in TensorFlow

Descriptions

This project is a TensorFlow implementation of a simple novel view synthesis model, which aims to synthesize a target view with an arbitrary camera pose from a given source view and its camera pose. An illustration of the task is as follows.

The model implemented in the repo is a simple conv-deconv network with skip connections. To allow you to focus on building your own model and see how well it can work, all the data loaders/downloading scripts, the training code, as well as the training and testing splits are well-configured based on the setting used in this paper: Multi-view to Novel view: Synthesizing Novel Views with Self-Learned Confidence published in ECCV 2018. All you need to do is play with the model code: synthesizer.py.

The model can be trained on images rendered from 3D object models (ShapeNet) as well as real and synthesized scenes (KITTI and Synthia). All datasets are stored as HDF5 files, they will be downloaded once you run the training code.

• ShapeNet: cars (150GB) and chairs (14GB)

• KITTI (4.3GB)

• Synthia (3.3GB)

Prerequisites

• Python 3.5
• Tensorflow 1.12.0
• NumPy
• colorlog
• h5py
• imageio
• six

Usage

After downloading the datasets, we can start to train and test models using the following commands

Train

Train a model from scratch

$ python trainer.py --batch_size 32 --dataset car

Fine-tune a model from a checkpoint

$ python trainer.py --batch_size 32 --dataset car --checkpoint /path/to/model/model-XXX

• Selected arguments (see the trainer.py for more details)
  • --prefix: a nickname for the training
  • --dataset: choose among car, chair, kitti, and synthia. You can also add your own datasets.
  • Checkpoints: specify the path to a pre-trained checkpoint
    • --checkpoint: load all the parameters including the flow and pixel modules and the discriminator.
  • Logging
    • --log_setp: the frequency of outputing log info ([train step 681] Loss: 0.51319 (1.896 sec/batch, 16.878 instances/sec))
    • --ckpt_save_step: the frequency of saving a checkpoint
    • --test_sample_step: the frequency of performing testing inference during training (default 100)
    • --write_summary_step: the frequency of writing TensorBoard summaries (default 100)
  • Hyperparameters
    • --batch_size: the mini-batch size (default 8)
    • --max_steps: the max training iterations

Interpret TensorBoard

Launch Tensorboard and go to the specified port, you can see differernt losses in the scalars tab and plotted images in the images tab. The scalars include L1 loss and SSIM. The plotted images show (from top to bottom): source view, target view, and the predicted target view.

Test

Evaluate trained models

$ python evaler.py --dataset car --loss True --plot_image True --output_dir car_image --write_summary True --summary_file log_car.txt --train_dir train_dir/default-car-bs_16_lr_0.0001-num_input-1-20190430-014454 --data_id_list ./testing_tuple_lists/id_car_random_elevation.txt

• Selected arguments (see the evaler.py for more details)

  • Id list
    • --data_id_list: specify a list of data point that you want to evaluate
  • Task
    • --loss: report the loss
    • --write_summary: write the summary of this evaluation as a text file
    • --plot_image: render synthesized images
  • Output
    • --quiet: only display the final report
    • --summary_file: the path to the summary file
    • --output_dir: the output dir of plotted images

• Synthesized images After plotting the synthesized images at the specified output_dir, we can see the results like this

• Loss summary The overall L1 loss and SSIM can be found at log_car.txt

Checkpoint: train_dir/default-car-bs_16_lr_0.0001-20190430-014454/model-160000
Dataset: car
Id list: ./testing_tuple_lists/id_car_elevation_0.txt
[Final Avg Report] Total datapoint: 10000 from ./testing_tuple_lists/id_car_elevation_0.txt
[Loss]
l1_loss: 0.13343
ssim: 0.90811
[Time] (63.128 sec)

Related work

The code is mainly borrowed from this paper

• Multi-view to Novel view: Synthesizing Novel Views with Self-Learned Confidence in ECCV 2018

Check out some other work in novel view synthesis

• Multi-view 3D Models from Single Images with a Convolutional Network in CVPR 2016
• View Synthesis by Appearance Flow in ECCV 2016
• Transformation-Grounded Image Generation Network for Novel 3D View Synthesis in CVPR 2017
• Neural scene representation and rendering in Science 2018
• Weakly-supervised Disentangling with Recurrent Transformations for 3D View Synthesis in NIPS 2015
• DeepStereo: Learning to Predict New Views From the World's Imagery in CVPR 2016
• Learning-Based View Synthesis for Light Field Cameras in SIGGRAPH Asia 2016

Cite the paper

If you find this useful, please cite

@inproceedings{sun2018multiview,
  title={Multi-view to Novel View: Synthesizing Novel Views with Self-Learned Confidence},
  author={Sun, Shao-Hua and Huh, Minyoung and Liao, Yuan-Hong and Zhang, Ning and Lim, Joseph J},
  booktitle={European Conference on Computer Vision},
  year={2018},
}

Authors

Shao-Hua Sun